diff options
| author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2022-08-25 12:25:44 +0100 |
|---|---|---|
| committer | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2022-09-02 08:33:21 +0000 |
| commit | 1257131193fdb9b6940055a41691320e37a208b5 (patch) | |
| tree | ea48d445c6be8e258d742a67fe41a65472d51963 | |
| parent | 16789a14afc27c1a77c8ca1e3d04b79cda6c833b (diff) | |
| download | ComputeLibrary-1257131193fdb9b6940055a41691320e37a208b5.tar.gz | |
Enable Winograd-based conv2d when IFM>=8 on Gpu
From an internal performance evaluation, it seems that Winograd-based
Conv2D offers better performance than alternative methods such as direct
convolution and gemm-based conv already from IFM=8. Before the condition
was for IFM>=16
Resolves COMPMID-5532
Change-Id: I9ff04835d6fd07f5f0abeec9645c9d9cc913b6b7
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8147
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
| -rw-r--r-- | src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp | 2 | ||||
| -rw-r--r-- | src/gpu/cl/operators/ClConv2d.cpp | 5 |
2 files changed, 4 insertions, 3 deletions
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp index d1f0338739..732d768308 100644 --- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp @@ -211,7 +211,7 @@ void CLDepthwiseConvolutionLayerNativeKernel::configure(const CLCompileContext & arm_compute::opencl::kernels::gemm::update_padding_for_cl_image(weights->info()); } - // Conditions of -cl-fast-relaxed-math causing accuracy issues can be traced from COMPMID-5324 + // Conditions of -cl-fast-relaxed-math causing accuracy issues can be traced from COMPMID-5324 const GPUTarget gpu_target = get_target(); const auto act_function = conv_info.act_info.activation(); const auto dst_data_type = _output->info()->data_type(); diff --git a/src/gpu/cl/operators/ClConv2d.cpp b/src/gpu/cl/operators/ClConv2d.cpp index 8119fc8e3d..16fc0e90d3 100644 --- a/src/gpu/cl/operators/ClConv2d.cpp +++ b/src/gpu/cl/operators/ClConv2d.cpp @@ -258,14 +258,15 @@ ConvolutionMethod ClConv2d::get_convolution_method(const ITensorInfo *src, const // Get dst shape TensorShape output_shape = misc::shape_calculator::compute_deep_convolution_shape(*src, *weights, conv_info); const bool is_large_kernel_sz = (weights->dimension(idx_w) >= kernel_sz_direct_conv_thr) && (weights->dimension(idx_h) >= kernel_sz_direct_conv_thr); + const bool is_ifm_ge_8 = src->dimension(idx_c) >= 8; const bool is_ifm_ge_16 = src->dimension(idx_c) >= 16; const bool is_ofm_lte_8 = weights->dimension(3U) <= 8; const bool workload_gte_8192 = (output_shape[0] * output_shape[1] * output_shape[2]) / 16 >= 8192; const bool is_ifm_gt_ofm = src->dimension(idx_c) > weights->dimension(3U); const bool is_m_one = output_shape[1] * output_shape[2] == 1; - // Run Winograd if valid and IFM >= 16 - if(is_wino_valid && is_ifm_ge_16) + // Run Winograd if valid and IFM >= 8 + if(is_wino_valid && is_ifm_ge_8) { if(is_ofm_lte_8) { |